MIT’s Senseable City Lab has used machine learning to classify vehicles seen by 331 New York City traffic cameras and estimate their emissions. The example, described in MIT News on September 24, illustrates how existing imagery can support urban research without treating every photograph as a separate manual counting exercise.
The researchers describe broader opportunities to study traffic, intersections and the use of public spaces. They also identify privacy and fairness as concerns when image analysis expands across a city. The emissions example involves estimates derived from observed vehicles; it should not be read as a direct measurement of exhaust from every automobile.
NAOAT analysis: Start with the decision
A useful pilot should begin with a specific planning question. Comparing traffic patterns before and after an intersection change calls for a different collection plan than estimating an annual emissions inventory. Deciding what action the results might support makes it easier to specify the necessary accuracy.
Coverage deserves its own evaluation. A camera network may observe major roads more frequently than side streets. Analysts should document those gaps before comparing neighborhoods, and check a sample of classifications against independent observations. A precise-looking map cannot compensate for a biased sample.
Procurement teams should also ask whether raw footage needs to be retained at all. Aggregate counts may answer the planning question without preserving identifiable journeys. Set retention periods and access rules alongside model validation, then publish enough methodology for residents to understand what the measurements can—and cannot—establish.